A Data-Driven Distributionally Robust Game Using Wasserstein Distance

Abstract

This paper studies a special class of games, which enables the players to leverage the information from a dataset to play the game. However, in an adversarial scenario, the dataset may not be trustworthy. We propose a distributionally robust formulation to introduce robustness against the worst-case scenario and tackle the curse of the optimizer. By applying Wasserstein distance as the distribution metric, we show that the game considered in this work is a generalization of the robust game and data-driven empirical game. We also show that as the number of data points in the dataset goes to infinity, the game considered in this work boils down to a Nash game. Moreover, we present the proof of the existence of distributionally robust equilibria and a tractable mathematical programming approach to solve for such equilibria.